Oh what you don't understand is that LLMs also use embeddings inside, it's how they represent tokens. It's just that you don't get to see the embeddings, they are inner workings.